MrBesen/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

overlay_vulkan: port for the rewrite

Browse Source
This commit is contained in:
Lynne 2023-02-17 03:13:05 +01:00
parent 59f21c051d
commit 8e9ceb1efb
No known key found for this signature in database
GPG Key ID: A2FEA5F03F034464
1 changed files with 112 additions and 264 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

376
libavfilter/vf_overlay_vulkan.c
View File

@ -1,4 +1,6 @@
/*
* Copyright (c) Lynne
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
@ -19,26 +21,26 @@
#include "libavutil/random_seed.h"
#include "libavutil/opt.h"
#include "vulkan_filter.h"
#include "vulkan_spirv.h"
#include "internal.h"
#include "framesync.h"
#define CGROUPS (int [3]){ 32, 32, 1 }
typedef struct OverlayVulkanContext {
FFVulkanContext vkctx;
FFFrameSync fs;
int initialized;
FFVulkanPipeline pl;
FFVkExecPool e;
FFVkQueueFamilyCtx qf;
FFVkExecContext *exec;
FFVulkanPipeline *pl;
FFFrameSync fs;
FFVkBuffer params_buf;
FFVkSPIRVShader shd;
VkSampler sampler;
/* Shader updators, must be in the main filter struct */
VkDescriptorImageInfo main_images[3];
VkDescriptorImageInfo overlay_images[3];
VkDescriptorImageInfo output_images[3];
VkDescriptorBufferInfo params_desc;
/* Push constants / options */
struct {
int32_t o_offset[2*3];
int32_t o_size[2*3];
} opts;
int overlay_x;
int overlay_y;
@ -80,279 +82,114 @@ static const char overlay_alpha[] = {
static av_cold int init_filter(AVFilterContext *ctx)
{
int err;
FFVkSampler *sampler;
uint8_t *spv_data;
size_t spv_len;
void *spv_opaque = NULL;
OverlayVulkanContext *s = ctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
const int ialpha = av_pix_fmt_desc_get(s->vkctx.input_format)->flags & AV_PIX_FMT_FLAG_ALPHA;
const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(s->vkctx.output_format);
FFVkSPIRVShader *shd = &s->shd;
FFVkSPIRVCompiler *spv;
FFVulkanDescriptorSetBinding *desc;
ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0);
sampler = ff_vk_init_sampler(vkctx, 1, VK_FILTER_NEAREST);
if (!sampler)
spv = ff_vk_spirv_init();
if (!spv) {
av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n");
return AVERROR_EXTERNAL;
s->pl = ff_vk_create_pipeline(vkctx, &s->qf);
if (!s->pl)
return AVERROR(ENOMEM);
{ /* Create the shader */
const int ialpha = av_pix_fmt_desc_get(s->vkctx.input_format)->flags & AV_PIX_FMT_FLAG_ALPHA;
FFVulkanDescriptorSetBinding desc_i[3] = {
{
.name = "main_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = s->main_images,
.sampler = sampler,
},
{
.name = "overlay_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = s->overlay_images,
.sampler = sampler,
},
{
.name = "output_img",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
.mem_quali = "writeonly",
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = s->output_images,
},
};
FFVulkanDescriptorSetBinding desc_b = {
.name = "params",
.type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
.mem_quali = "readonly",
.mem_layout = "std430",
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = &s->params_desc,
.buf_content = "ivec2 o_offset[3], o_size[3];",
};
FFVkSPIRVShader *shd = ff_vk_init_shader(s->pl, "overlay_compute",
VK_SHADER_STAGE_COMPUTE_BIT);
if (!shd)
return AVERROR(ENOMEM);
ff_vk_set_compute_shader_sizes(shd, CGROUPS);
RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, desc_i, FF_ARRAY_ELEMS(desc_i), 0)); /* set 0 */
RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, &desc_b, 1, 0)); /* set 1 */
GLSLD( overlay_noalpha );
GLSLD( overlay_alpha );
GLSLC(0, void main() );
GLSLC(0, { );
GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
GLSLF(1, int planes = %i; ,planes);
GLSLC(1, for (int i = 0; i < planes; i++) { );
if (ialpha)
GLSLC(2, overlay_alpha_opaque(i, pos); );
else
GLSLC(2, overlay_noalpha(i, pos); );
GLSLC(1, } );
GLSLC(0, } );
RET(ff_vk_compile_shader(vkctx, shd, "main"));
}
RET(ff_vk_init_pipeline_layout(vkctx, s->pl));
RET(ff_vk_init_compute_pipeline(vkctx, s->pl));
ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT);
RET(ff_vk_exec_pool_init(vkctx, &s->qf, &s->e, s->qf.nb_queues*4, 0, 0, 0, NULL));
RET(ff_vk_init_sampler(vkctx, &s->sampler, 1, VK_FILTER_NEAREST));
RET(ff_vk_shader_init(&s->pl, &s->shd, "overlay_compute",
VK_SHADER_STAGE_COMPUTE_BIT, 0));
{ /* Create and update buffer */
const AVPixFmtDescriptor *desc;
ff_vk_shader_set_compute_sizes(&s->shd, 32, 32, 1);
/* NOTE: std430 requires the same identical struct layout, padding and
* alignment as C, so we're allowed to do this, as this will map
* exactly to what the shader recieves */
struct {
int32_t o_offset[2*3];
int32_t o_size[2*3];
} *par;
GLSLC(0, layout(push_constant, std430) uniform pushConstants { );
GLSLC(1, ivec2 o_offset[3]; );
GLSLC(1, ivec2 o_size[3]; );
GLSLC(0, }; );
GLSLC(0, );
err = ff_vk_create_buf(vkctx, &s->params_buf,
sizeof(*par), NULL,
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
if (err)
return err;
ff_vk_add_push_constant(&s->pl, 0, sizeof(s->opts),
VK_SHADER_STAGE_COMPUTE_BIT);
err = ff_vk_map_buffers(vkctx, &s->params_buf, (uint8_t **)&par, 1, 0);
if (err)
return err;
desc = (FFVulkanDescriptorSetBinding []) {
{
.name = "main_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(s->sampler),
},
{
.name = "overlay_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(s->sampler),
},
{
.name = "output_img",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
.mem_quali = "writeonly",
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
},
};
desc = av_pix_fmt_desc_get(s->vkctx.output_format);
RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl, shd, desc, 3, 0, 0));
par->o_offset[0] = s->overlay_x;
par->o_offset[1] = s->overlay_y;
par->o_offset[2] = par->o_offset[0] >> desc->log2_chroma_w;
par->o_offset[3] = par->o_offset[1] >> desc->log2_chroma_h;
par->o_offset[4] = par->o_offset[0] >> desc->log2_chroma_w;
par->o_offset[5] = par->o_offset[1] >> desc->log2_chroma_h;
GLSLD( overlay_noalpha );
GLSLD( overlay_alpha );
GLSLC(0, void main() );
GLSLC(0, { );
GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
GLSLF(1, int planes = %i; ,planes);
GLSLC(1, for (int i = 0; i < planes; i++) { );
if (ialpha)
GLSLC(2, overlay_alpha_opaque(i, pos); );
else
GLSLC(2, overlay_noalpha(i, pos); );
GLSLC(1, } );
GLSLC(0, } );
par->o_size[0] = s->overlay_w;
par->o_size[1] = s->overlay_h;
par->o_size[2] = par->o_size[0] >> desc->log2_chroma_w;
par->o_size[3] = par->o_size[1] >> desc->log2_chroma_h;
par->o_size[4] = par->o_size[0] >> desc->log2_chroma_w;
par->o_size[5] = par->o_size[1] >> desc->log2_chroma_h;
RET(spv->compile_shader(spv, ctx, shd, &spv_data, &spv_len, "main",
&spv_opaque));
RET(ff_vk_shader_create(vkctx, shd, spv_data, spv_len, "main"));
err = ff_vk_unmap_buffers(vkctx, &s->params_buf, 1, 1);
if (err)
return err;
RET(ff_vk_init_compute_pipeline(vkctx, &s->pl, shd));
RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl));
s->params_desc.buffer = s->params_buf.buf;
s->params_desc.range = VK_WHOLE_SIZE;
s->opts.o_offset[0] = s->overlay_x;
s->opts.o_offset[1] = s->overlay_y;
s->opts.o_offset[2] = s->opts.o_offset[0] >> pix_desc->log2_chroma_w;
s->opts.o_offset[3] = s->opts.o_offset[1] >> pix_desc->log2_chroma_h;
s->opts.o_offset[4] = s->opts.o_offset[0] >> pix_desc->log2_chroma_w;
s->opts.o_offset[5] = s->opts.o_offset[1] >> pix_desc->log2_chroma_h;
ff_vk_update_descriptor_set(vkctx, s->pl, 1);
}
/* Execution context */
RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf));
s->opts.o_size[0] = s->overlay_w;
s->opts.o_size[1] = s->overlay_h;
s->opts.o_size[2] = s->opts.o_size[0] >> pix_desc->log2_chroma_w;
s->opts.o_size[3] = s->opts.o_size[1] >> pix_desc->log2_chroma_h;
s->opts.o_size[4] = s->opts.o_size[0] >> pix_desc->log2_chroma_w;
s->opts.o_size[5] = s->opts.o_size[1] >> pix_desc->log2_chroma_h;
s->initialized = 1;
return 0;
fail:
return err;
}
if (spv_opaque)
spv->free_shader(spv, &spv_opaque);
if (spv)
spv->uninit(&spv);
static int process_frames(AVFilterContext *avctx, AVFrame *out_f,
AVFrame *main_f, AVFrame *overlay_f)
{
int err;
VkCommandBuffer cmd_buf;
OverlayVulkanContext *s = avctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
FFVulkanFunctions *vk = &vkctx->vkfn;
int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
AVVkFrame *out = (AVVkFrame *)out_f->data[0];
AVVkFrame *main = (AVVkFrame *)main_f->data[0];
AVVkFrame *overlay = (AVVkFrame *)overlay_f->data[0];
AVHWFramesContext *main_fc = (AVHWFramesContext*)main_f->hw_frames_ctx->data;
AVHWFramesContext *overlay_fc = (AVHWFramesContext*)overlay_f->hw_frames_ctx->data;
const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format);
const VkFormat *main_sw_formats = av_vkfmt_from_pixfmt(main_fc->sw_format);
const VkFormat *overlay_sw_formats = av_vkfmt_from_pixfmt(overlay_fc->sw_format);
/* Update descriptors and init the exec context */
ff_vk_start_exec_recording(vkctx, s->exec);
cmd_buf = ff_vk_get_exec_buf(s->exec);
for (int i = 0; i < planes; i++) {
RET(ff_vk_create_imageview(vkctx, s->exec,
&s->main_images[i].imageView, main->img[i],
main_sw_formats[i],
ff_comp_identity_map));
RET(ff_vk_create_imageview(vkctx, s->exec,
&s->overlay_images[i].imageView, overlay->img[i],
overlay_sw_formats[i],
ff_comp_identity_map));
RET(ff_vk_create_imageview(vkctx, s->exec,
&s->output_images[i].imageView, out->img[i],
output_formats[i],
ff_comp_identity_map));
s->main_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
s->overlay_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
}
ff_vk_update_descriptor_set(vkctx, s->pl, 0);
for (int i = 0; i < planes; i++) {
VkImageMemoryBarrier bar[3] = {
{
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.srcAccessMask = 0,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
.oldLayout = main->layout[i],
.newLayout = s->main_images[i].imageLayout,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = main->img[i],
.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.subresourceRange.levelCount = 1,
.subresourceRange.layerCount = 1,
},
{
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.srcAccessMask = 0,
.dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
.oldLayout = overlay->layout[i],
.newLayout = s->overlay_images[i].imageLayout,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = overlay->img[i],
.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.subresourceRange.levelCount = 1,
.subresourceRange.layerCount = 1,
},
{
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.srcAccessMask = 0,
.dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
.oldLayout = out->layout[i],
.newLayout = s->output_images[i].imageLayout,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = out->img[i],
.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.subresourceRange.levelCount = 1,
.subresourceRange.layerCount = 1,
},
};
vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
0, NULL, 0, NULL, FF_ARRAY_ELEMS(bar), bar);
main->layout[i] = bar[0].newLayout;
main->access[i] = bar[0].dstAccessMask;
overlay->layout[i] = bar[1].newLayout;
overlay->access[i] = bar[1].dstAccessMask;
out->layout[i] = bar[2].newLayout;
out->access[i] = bar[2].dstAccessMask;
}
ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl);
vk->CmdDispatch(cmd_buf,
FFALIGN(s->vkctx.output_width, CGROUPS[0])/CGROUPS[0],
FFALIGN(s->vkctx.output_height, CGROUPS[1])/CGROUPS[1], 1);
ff_vk_add_exec_dep(vkctx, s->exec, main_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
ff_vk_add_exec_dep(vkctx, s->exec, overlay_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
ff_vk_add_exec_dep(vkctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
err = ff_vk_submit_exec_queue(vkctx, s->exec);
if (err)
return err;
ff_vk_qf_rotate(&s->qf);
return err;
fail:
ff_vk_discard_exec_deps(s->exec);
return err;
}
@ -394,7 +231,9 @@ static int overlay_vulkan_blend(FFFrameSync *fs)
goto fail;
}
RET(process_frames(ctx, out, input_main, input_overlay));
RET(ff_vk_filter_process_Nin(&s->vkctx, &s->e, &s->pl,
out, (AVFrame *[]){ input_main, input_overlay }, 2,
s->sampler, &s->opts, sizeof(s->opts)));
err = av_frame_copy_props(out, input_main);
if (err < 0)
@ -443,8 +282,17 @@ static av_cold int overlay_vulkan_init(AVFilterContext *avctx)
static void overlay_vulkan_uninit(AVFilterContext *avctx)
{
OverlayVulkanContext *s = avctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
FFVulkanFunctions *vk = &vkctx->vkfn;
ff_vk_exec_pool_free(vkctx, &s->e);
ff_vk_pipeline_free(vkctx, &s->pl);
ff_vk_shader_free(vkctx, &s->shd);
if (s->sampler)
vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler,
vkctx->hwctx->alloc);
ff_vk_free_buf(&s->vkctx, &s->params_buf);
ff_vk_uninit(&s->vkctx);
ff_framesync_uninit(&s->fs);